noticias ia para empresas microsoft copilot agentes de ia

AI billing is going on the meter: Microsoft will activate pay-per-use by default on December 1st for Copilot Business, with a cap of 4,000 credits per user per month, and Google is launching Gemini 4 Argon with an introductory price that will later double. Infrastructure money is now moving in circles: Broadcom will lend up to $42 billion to Anthropic so it can lease its chips. And free sources are closing: Reddit is turning off its RSS feeds on November 13th and its public API in March. This follows Thursday's radar (October 1).

Microsoft to activate pay-per-use for Copilot Business on December 1st: 4,000 credits per user/month, unless you set the cap

Microsoft notified its partners yesterday that, starting December 1st (it had previously said November 2)new Microsoft 365 Copilot Business licenses purchased through a partner or distributor (the CSP channel), whether standalone or bundled, will come with pay-per-use activated by default. This meter tracks features that are charged separately, such as Copilot Cowork, Work IQ APIs, and the GitHub Copilot harness. The default spending limit will be 4,000 credits per user per month, adjustable by the administrator: using Microsoft's own example, a 100-user company could consume up to 400,000 credits per month. Microsoft sells these credits in packs of 25,000 for $200; at that price, the default cap comes to about $32 per user, more than the license itself ($21 list price in the US, €18.20 in Spain). Copilot Business is Microsoft's plan for companies with up to 300 users: the SMB segment. For your business: in the September 25th flash, we mentioned that the part of Copilot that actually does the heavy lifting doesn't have a fixed price; now we know that tap will be open by default. Three things to do before December in the Microsoft 365 cost panel: lower the limit to zero for everyone except the group testing Cowork, with a written monthly budget; disable the option that automatically adds new services and agents to your spending policies (which, according to Microsoft documentation, is enabled by default); and review consumption every week during the first month. Also, ask your partner in writing what limit they will set for your new licenses. An agent that works alone and without a spending cap isn't UNATTENDED: it's ungoverned. Source

Gemini 4 Argon: Google's most powerful model scores 51.3% on business tasks, cannot be purchased yet, and price will double

On Wednesday night, Google unveiled Gemini 4 Argon, its new frontier model. For now it's going only to Google's own teams and to trusted cyber defenders in its Fairwind Program, and they get it without cyber guardrails. According to Google, it leads the Vals Index, which measures work in finance, programming, law, and taxes, and is number one in AutomationBench, Zapier's benchmark for whether a model can carry out everyday business tasks end to end, with a score of 51.3%. Wiz is already using it to scan critical infrastructure, and with it found a critical flaw exposing personal data in healthcare software used by hospitals worldwide. It will arrive first for paid API customers and Google AI Ultra subscribers, with no date yet. The introductory price is $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20: double, and Google hasn't said when. For your company, two takeaways. If the best model of the moment, according to its maker, barely clears half on office tasks, the rule still stands: a process only becomes UNATTENDED where there is an automatic check of the result (does the ledger entry balance? are all the fields there?); where there isn't, the AI prepares and a human approves. Then there's the price: as with Gemini 3.8 Flash on September 3rd, a launch rate is a promotion, not a price, and this time it doesn't even have an expiration date. If you're running numbers for a project with Argon, run them at $4 and $20. Source

Broadcom to lend up to $42 billion to Anthropic to lease its chips: when the one who sells to you also finances you

Broadcom has committed to lending Anthropic up to $42 billion in convertible notes to fund its infrastructure, according to Anthropic's IPO prospectus reviewed by Reuters. That money would cover about a third of the $125.2 billion Anthropic has committed to pay over five years for compute capacity on TPUs, Google's chips, which Broadcom helps build; the notes can convert into Anthropic shares, and none will be sold before the IPO. Broadcom thus provides the chips, the lease, and the loan. The prospectus itself warns of "potential conflicts of interest": Broadcom's pricing and hardware decisions could affect Anthropic's ability to get the compute it needs. "It feels that there's quite a concentrated bet right now on two companies being able to generate enough revenues to support all the financing that's happened," Robert Leitao, managing partner at Rothschild & Co, told Reuters. From the prospectus we covered on September 30th, with over $518 billion committed to six partners, we now learn that one of them is also a lender. For your company: the same scheme appears on a smaller scale when the integrator building your AI also finances it—leasing, deferred payment through their finance arm, a three-year license with a discount—and from that day on, you negotiate price and exit with someone who is also your creditor. Before signing, separate the service from the financing in the contract and ask for three answers in writing: if you can terminate the service without canceling the loan, what comes due early if you miss a payment, and who owns the configurations and flows at the end. AI money isn't just flowing in anymore: it's going in circles between seller and buyer. Source

Reddit closes its RSS on November 13th and public API in March: every automation reading a free source has an expiration date

Reddit will stop serving its RSS feeds on November 13th and will close public access to its API in March 2027; approved third-party apps and bots must register by January 12th or lose access, and social listening tools and AI assistants reading Reddit will need a commercial agreement. Its argument: RSS has become a "common surface for large-scale scraping and automated abuse." Meanwhile, its revenue outside of advertising—mostly data licensing for AI—grew 24% in the second quarter to $43 million. For your company: if you monitor mentions of your brand, competition, or sector on Reddit via RSS in Make, Zapier, or n8n, the feed goes dark on November 13th; and if you use a social listening tool, ask your provider if they have an agreement or registered app. On September 16th it was Cloudflare blocking crawlers: open data is becoming paid data. Make a list of your automations that read free sources—RSS, APIs without contracts, scraped pages—and set a silence alarm for each: if it doesn't bring data for two consecutive days, have it notify a person. An UNATTENDED process reading an empty channel doesn't fail: it just stays quiet. Source

What to watch tomorrow?

Whether Microsoft spells out what each Copilot Business credit costs under pay-per-use, and whether the change also covers renewals, not just new licenses. And whether Google sets a date for the Argon paid API and the end of its introductory pricing.